Summary

  • Together AI and Agentica have developed an AI model called DeepCoder-14B, offering competitive coding and mathematical capabilities, which they claim is on a par with Open AI’s o3-mini.
  • Built on top of previous AI software DeepSeek-R1, DeepCoder is completely open source, meaning all data, training, code, logs and optimisation information is available to the public.
  • The developers say this is to help fellow researchers to enhance and accelerate their work in the field.
  • It also performs well on mathematical tests, showing a 4.1% improvement on its predecessor, demonstrating that the skills it has learned through reinforcement learning can be applied to other areas.
  • The makers of DeepCoder have shared their training data and code, as well as the system they used, on GitHub and Huggingface with a view to empowering the coding community with the skills to reproduce, develop and enhance the model.

By Ben Dickson

Original Article